A deep learning framework for automated large-area detection of individual tree crowns and classification of tree species (ML4iTree =Machine Learning for identification of Tree species)

• Applicants: Professor Dr. Alexander Ecker; Professor Dr. Christoph Kleinn
• Subject Area: Forestry
• Term: since 2022
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 494541002

Forests in Central Europe undergo drastic changes, manifesting themselves in (1) gradual changes in site conditions (changing rainfall and temperature patterns, atmospheric intakes) and (2) an increasing number of abiotic and biotic disasters. Forest research is searching for options how to adapt forest and ecosystem management to these unprecedented challenges. Accurate and timely information on status and changes of the forests constitute an indispen-sable precondition for such adaptation. However, there are manifold methodological challen-ges concerning the monitoring of individual trees over large areas for which high-resolution remote sensing imagery has become more and more available and offers corresponding analysis options.The big data involved when working with high-resolution imagery on larger areas make conventional classification approaches fail. In this project, we will adapt and apply deep learning techniques ‑ notably convolutional neural networks ‑ for large-area forest monitoring to simultaneously (1) identify individual trees and (2) identify their species. So far, this is commonly done in two steps with two different remote sensing data sources. We plan to work with only one data source: high-resolution optical imagery, without the need to analyse lidar data first which are costly for larger areas. From own available and additional optical image data sets we will generate (1) a large archive of unlabelled image patches and (2) with extensive ground-truthing training a training data set on tree species for Central European tree species. The archive can gradually be enhanced and will be made available to other research teams as a basis for further deep learning applications in forestry. We will develop a methods´ toolbox for large area mapping of individual trees’ species that may be applied to whole countries, given availability of the corresponding imagery.We join forces between the two methodologically oriented disciplines “forest monitoring science” (PI Kleinn) and “data science” (PI Ecker) - and implicitly respond to current policies in BMEL (request to further develop efficiency of forest condition monitoring) and BMBF (fostering applications of artificial intelligence).Our research will provide the methodological basis for more efficient monitoring systems and thus contribute to the adaptation of forest ecosystem management to climate change; it generates a showcase for application of deep learning for forest monitoring.

• DFG Programme: Research Grants